GUIDED MCMC FOR SPARSE BAYESIAN MODELS TO DETECT RARE EVENTS IN IMAGES SANS LABELED DATA

Gaurav Jain, Mrinal Das

Keywords: MCMC, hierarchical Bayesian models, image classification, rare events

Abstract: Detection of rare events in images is a challenging task because of two main problems, the first problem is the lack of labeled data for rare category class and the second problem is a highly imbalanced data problem. Training models in this scenario becomes hard. Unsupervised methods do not apply as we need to detect rare events automatically. Rule-based methods seem to be the only viable solution, but it is tedious to come up with a set of rules covering all corner cases. Even the recently popular zero-shot learning techniques required to be pre-trained on auxiliary datasets. In the given scenario, we propose an approach to provide little guidance from experts as an input into a hierarchical Bayesian model. The guidance influences the Markov chain Monte Carlo (MCMC) based inference technique of the model. After the steady-state is obtained for the underlying Markov chain, it is possible to compute the posterior probability of the presence of the rare event in a given image. The proposed method neither needs any labeled data nor required pre-training, unlike zero-shot learning. The proposed technique has been observed to outperform the state-of-the-art unsupervised image classification techniques.